|
Boost : |
From: Petr Ovchenkov (ptr_at_[hidden])
Date: 2002-01-04 06:36:23
Hi,
For number of dimentions there are some very important cases,
when performance is a key issue:
N = 2
N = 3
N = 4 (may be)
The performance for this dimentions can be near (or better) then
FORTRAN code^# if we roll out many internal loops with recursive
templates, and, may be, with specialized algorithms.
---------------------------------------------
^# [as shown by Todd Veldhuizen in Blitz++]
The absent of specialization for this two-three dimentions in
multiarray librarias (as mtl, as ublas) is a significant drawback,
IMO.
I think, that interfaces of specialization for N=2,..,4 should be
the same (or very similar) as for common case.
- ptr
Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk